Underdetermined Blind Speech Separation Based On Time-frequency Sparsity

Blind sources separation is process of separating the source signal from a set of unknown random signals mixed signals, if we do not need any priori knowledge of the source signal and its channel, this separation process is called blind source separation. Most of blind sources algorithms assume that the source number is known and the number of mixed signals is equal or more than the number of sources. In practice, the source number is unknown and the number of mixed signals is less than the number of sources. In this dissertation, we focus on investigating the problem of the underdetermined algorithm. The main contributions of this dissertation are summarized as follows:Under the framework of underdetermined blind anechoic speech separation based on time-frequency sparsity, we propose a improved method for the separation of speech mixtures using two-microphone recordings based on combination of EFICA algorithm and binary-mask, together with a post-filtering process in the cepstral domain when the number of sources is unknown. The accuracy of the Efficient Variant of FastICA(EFICA) algorithm given by the residual error variance attains the Cramer-Rao lower bound. The error is thus as small as possible and adaptive nonlinear function is estimated. The cepstral smoothing operation is to reduce musical noise caused by T-F masking. The proposed algorithm improved speech separation quality measured by the Percentage of Energy Loss(PEL), the Signal-to-Noise Ratio(SNR), the Signal to Distortion Ratio(SDR) and the Sources to Artifacts Ratio(SAR).